Project Description

A production-ready RAG application that serves as an intelligent 'Digital Twin,' answering queries about my professional history using a hybrid retrieval pipeline and serverless state management.

Responsibilities

• Architected a serverless RAG pipeline on Google Cloud Run, utilizing LlamaIndex's 'CondensePlusContextChatEngine' to maintain deep conversational context across multiple turns.
• Engineered a robust state management system using Firestore to persist chat history server-side, enabling both long-term memory for the user and offline quality analysis for the developer.
• Optimized Docker container sizing by over 70% (from 5.5GB to ~1.5GB) by enforcing CPU-only PyTorch wheels and implementing a multi-stage build process that excludes heavy CUDA GPU drivers.
• Solved the 'Serverless Cold Start' problem (typically 40s+ for LLM apps) by deploying a Cloud Scheduler 'Warmer Bot' that pings the service every 15 minutes, ensuring sub-second response times.
• Implemented a two-stage retrieval strategy: broad semantic search using 'all-MiniLM-L6-v2' followed by a precision pass using the 'ms-marco-MiniLM-L-2-v2' Cross-Encoder to minimize hallucination.
• Built a polished, terminal-themed UI using Next.js and Tailwind CSS, featuring a custom 'useSession' hook that generates and persists UUIDs client-side to maintain chat history across page reloads.
• Configured a production-grade 'autoscaling.knative.dev' strategy with a max-instance limit of 5, balancing high availability with strict cost controls.
• Integrated a 'Typewriter' effect and 'Suggested Query Chips' to guide user interaction and reduce the cognitive load of the 'Blank Page Problem' common in chatbot interfaces.

Related Links

• Chat with Me

Technology